What does it cost to add AI to an existing app?
Adding AI to an app means a one-off build plus a per-use bill. How to estimate the running cost from published model prices, where build effort really goes, and the mistakes that inflate both.
- Author
Umair Mateen- Published
- Reading time
- 6 min
- Category
- AI
Ideas in this post
Tap one for a 5-second answer. Underlined words in the article work the same way.
0 of 4 understood

Adding AI to an app you already have costs two things: a one-off build (connecting a model to your app, your data and your screens, then testing it) and a running bill for every request the feature makes. The running bill is usually the easier part to predict, because AI providers publish per-use prices. The build cost depends almost entirely on scope, so the cheapest way to control it is to start with one narrow feature.
This guide shows where the money goes, how to estimate the running bill yourself from published prices, and the mistakes that make AI features cost more than planned.
Step 1: Name one feature, not "AI"
"Add AI" cannot be priced. "Draft a reply to each support email for an agent to approve" can. Write down one feature in one sentence: what goes in (an email, a scanned invoice, a question), what comes out (a draft, extracted fields, an answer), who checks it and how often it runs. That sentence decides almost every cost below.
Quick check · Put it in order
Put the steps of an AI eval in order
Tap the steps in the order they happen.
From the explainer What are AI evals? Explained with a driving test
Step 2: Understand the one-off build costs
Most of the build effort is not the AI itself. Calling a model is a few lines of code. The work sits around it:
- Connecting to your data. Pulling the right records, documents or help articles into each request, with the right permissions.
- Screens and workflow. Where the result appears, how a person approves or corrects it, and what happens when the AI is unsure.
- Testing for quality. A set of real examples with known good answers, run before launch and after every change. These are called AI evals.
- Safety and privacy. Deciding what data may be sent to the provider, logging, and limits on what the feature can do.
- Monitoring. Tracking cost per request, errors and user corrections once it is live.
There is no reliable public for what this build costs, because a feature that reads one clean is a very different job from one that pulls from five old systems. Treat any quoted figure without a written scope as a guess. Ask for a short, fixed-price that ends with a scope, an estimate and a test plan.
Step 3: Estimate the running bill from published prices
AI providers charge by tokens, small chunks of text (roughly a short word or part of a word). You pay for the tokens you send in (your instructions plus the user's content) and, at a higher rate, for the tokens the model writes back. Prices are quoted per million tokens. The four numbers you need are:
- Input tokens per request (instructions + content).
- Output tokens per request (the answer).
- Requests per month.
- The model's input and output prices.
As a worked example, take a feature that sends 2,000 input tokens and gets 500 output tokens back, 1,000 times a day. That is 60 million input tokens and 15 million output tokens over 30 days. Using prices published today on the Claude pricing page and the Gemini API pricing page (standard tier, US dollars per million tokens):
| Model | Input / output price | Monthly estimate |
|---|---|---|
| Claude Opus 5.5 | $4 / $20 | $540 |
| Claude Sonnet 5.5 | $2 / $10 | $270 |
| Gemini 3.5 Flash | $1.50 / $9.00 | $225 |
| Gemini 3.1 Flash-Lite (text) | $0.25 / $1.50 | $37.50 |
| Claude Haiku 5.5 (prompts up to 100,000 tokens) | $0.10 / $0.50 | $13.50 |
These are illustrations calculated from list prices, not quotes. Your real numbers depend on how long your prompts and answers are. Two details from the pricing pages change the maths: Google counts the model's "thinking" as output tokens, and Anthropic notes that its newer models use a tokenizer that produces about 30% more tokens for the same text. Measure token counts on your own examples before you trust any estimate.
Step 4: Use the built-in discounts
- Batch processing. Work that does not need an instant answer, such as overnight document tagging, can go through a batch API. Both Anthropic and Google list a 50% discount for it. See what batch processing is.
- Prompt caching. If every request repeats the same long instructions, the provider can store them. On Anthropic's page, reading cached input on Claude Sonnet 5.5 costs $0.10 per million tokens against $2 for normal input; writing to the cache costs more than normal input, so it pays off only when the same text is reused. See what prompt caching is.
- Smaller models for simple jobs. The table shows a gap of more than 30 times between the most and least expensive options. Classifying or extracting fields often works on a small model; test it with your evals before paying for a larger one.
Step 5: Budget for the extras
Some features pay for tools as well as tokens. For example, Google lists grounding answers with Google Search at 5,000 prompts per month free (shared across Gemini 3 models), then $14 per 1,000 search queries. Add hosting for any new parts of your app, monitoring, and time each month to review results and update prompts when models change.
What it costs and how long it takes
- Running cost: calculable from published per-token prices, as above. For the same workload, the example ranges from about $13.50 to $540 a month depending on the model.
- Build cost and timeline: set by scope, data access and testing, not by the model. No trustworthy public average exists; get a written scope first.
- Free tiers: useful for trying ideas, but check the terms. Google's pricing page says content on the Gemini free tier is used to improve its products, while paid-tier content is not, so keep customer data off free tiers.
Common mistakes
- Pricing "AI" in general instead of one feature with a clear input, output and owner.
- Estimating tokens from a guess instead of measuring real examples.
- Defaulting to the largest model for every task.
- Forgetting that a chat feature usually sends the earlier conversation again with each new message, so long chats cost more per message.
- Skipping a spending limit and usage alerts on the provider account.
- Budgeting for the build but not for testing, monitoring and monthly review.
A checklist you can use today
- Write the one feature you want in one sentence: input, output, who checks it.
- Estimate how many times a month it will run.
- Collect 20 to 30 real examples and the answers you would accept.
- Measure input and output tokens on those examples with two or three models.
- Multiply by current list prices from the provider's own pricing page.
- Mark which work can wait for batch processing and which instructions repeat.
- Confirm you will use a paid tier for any customer or company data.
- Set a monthly spending cap and an alert before launch.
- Ask any developer for a written scope and test plan before a build price.
If you want help scoping an AI feature for an app you already run and estimating its real running cost, The DevDrift can help: see our AI services.
Sources
You’ve read the news. Now understand it.
- Learn · 5 min · Put it in orderWhat are AI evals? Explained with a driving testAn AI eval is a set of real test tasks with clear marking rules that checks your AI system is good enough before it goes live, and after every change. Think of it as a driving test.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat are AI benchmarks? Explained with a school examAn AI benchmark is a standard test: the same questions and the same marking for every model, so scores can be compared. Explained with a school exam that every student sits.Start in 5 seconds →
- Learn · 5 min · Put it in orderWhat is a database? Explained with a filing cabinet and its indexA database is an organised store of information that software keeps tidy, searchable and safe for many users at once. Explained as an office filing cabinet with a perfect card index.Start in 5 seconds →
- Learn · 5 min · Myth or factWhat is project discovery? Explained with an architect's briefDiscovery is the short research stage before a software project is built: learning who it is for, the real problem and the limits. Explained with an architect's brief written before any plans.Start in 5 seconds →


